SIYI 相机 + YOLO + UDP 推流

系统架构

SIYI Camera RTSP ↓ Jetson / Orin Python OpenCV ↓ YOLO human detection ↓ FFmpeg encodes video ↓ UDP stream to Windows ↓ VLC opens udp://@0.0.0.0:5600

配置参数

参数	值	说明
RTSP_URL	`rtsp://192.168.144.25:8554/main.264`	SIYI 相机 RTSP 流地址
WINDOWS_IP	`192.168.1.14`	Windows 电脑 IP（UDP 接收端）
UDP_PORT	`5600`	UDP 推流端口
MODEL_PATH	`/home/drobotics/yolov8n.pt`	YOLOv8 模型路径
CONF_THRES	`0.25`	检测置信度阈值
FRAME_WIDTH	`640`	输出帧宽度
FRAME_HEIGHT	`480`	输出帧高度
FPS	`25`	输出帧率

完整 Python 脚本

#!/usr/bin/env python3

import cv2
import time
import subprocess
from ultralytics import YOLO

RTSP_URL = "rtsp://192.168.144.25:8554/main.264"
WINDOWS_IP = "192.168.1.14"
UDP_PORT = 5600

MODEL_PATH = "/home/drobotics/yolov8n.pt"
CONF_THRES = 0.25

FRAME_WIDTH = 640
FRAME_HEIGHT = 480
FPS = 25

def open_camera():
 cap = cv2.VideoCapture(RTSP_URL, cv2.CAP_FFMPEG)
 cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
 return cap

print("[SIYI YOLO] Loading model...")
model = YOLO(MODEL_PATH)

print("[SIYI YOLO] Opening camera...")
cap = open_camera()

if not cap.isOpened():
 print("[SIYI YOLO] ERROR: Cannot open RTSP stream")
 exit(1)

ffmpeg_cmd = [
 "ffmpeg",
 "-y",
 "-loglevel", "warning",
 "-fflags", "nobuffer",
 "-flags", "low_delay",
 "-f", "rawvideo",
 "-vcodec", "rawvideo",
 "-pix_fmt", "bgr24",
 "-s", f"{FRAME_WIDTH}x{FRAME_HEIGHT}",
 "-r", str(FPS),
 "-i", "-",
 "-an",
 "-c:v", "libx264",
 "-preset", "ultrafast",
 "-tune", "zerolatency",
 "-pix_fmt", "yuv420p",
 "-b:v", "2500k",
 "-maxrate", "2500k",
 "-bufsize", "500k",
 "-f", "mpegts",
 f"udp://{WINDOWS_IP}:{UDP_PORT}?pkt_size=1316"
]
print("[SIYI YOLO] Opening FFmpeg UDP output...")
process = subprocess.Popen(ffmpeg_cmd, stdin=subprocess.PIPE)

print("[SIYI YOLO] Running. Open VLC on Windows:")
print(f"udp://@:{UDP_PORT}")

frame_count = 0
last_time = time.time()

try:
 while True:
 ret, frame = cap.read()

 if not ret:
 print("[SIYI YOLO] No frame, reconnecting...")
 cap.release()
 time.sleep(1)
 cap = open_camera()
 continue

 frame = cv2.resize(frame, (FRAME_WIDTH, FRAME_HEIGHT))

 results = model(frame, conf=CONF_THRES, classes=[0], verbose=False)

 detection_count = 0

 for r in results:
 for box in r.boxes:
 detection_count += 1

 cls = int(box.cls[0])
 conf = float(box.conf[0])
 name = model.names[cls]

 x1, y1, x2, y2 = map(int, box.xyxy[0])

 cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
 cv2.putText(
 frame,
 f"{name} {conf:.2f}",
 (x1, max(y1 - 10, 20)),
 cv2.FONT_HERSHEY_SIMPLEX,
 0.6,
 (0, 255, 0),
 2
 )

 frame_count += 1

 if time.time() - last_time >= 1:
 print(f"[SIYI YOLO] FPS: {frame_count}, person detections: {detection_count}")
 frame_count = 0
 last_time = time.time()

 process.stdin.write(frame.tobytes())

except KeyboardInterrupt:
 print("\n[SIYI YOLO] Stopped by user")

except BrokenPipeError:
 print("[SIYI YOLO] ERROR: FFmpeg pipe broken")

finally:
 cap.release()
 if process.stdin:
 process.stdin.close()
 process.wait()
 print("[SIYI YOLO] Clean exit")

代码功能详解

1 配置参数

RTSP_URL：SIYI 相机 RTSP 流地址
WINDOWS_IP / UDP_PORT：UDP 推流目标
MODEL_PATH / CONF_THRES：YOLO 模型路径与置信度阈值
FRAME_WIDTH / FRAME_HEIGHT / FPS：输出视频参数

2 相机打开 (open_camera)

cap = cv2.VideoCapture(RTSP_URL, cv2.CAP_FFMPEG)
cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)

使用 FFmpeg 后端打开 RTSP 流，设置缓冲区大小为 1 以减少延迟。

3 FFmpeg 子进程

- 输入：原始 BGR24 帧 (640x480, 25fps)
- 编码：libx264 (ultrafast, zerolatency)
- 码率：2500k
- 输出：mpegts UDP 推流到 Windows

4 YOLO 检测循环

- 读取 RTSP 帧 → resize
- YOLO 推理 (classes=[0] 只检测人, conf=0.25)
- 绘制检测框和标签
- 帧率统计 (每秒输出一次)
- 帧数据写入 FFmpeg 子进程

5 断线重连

if not ret:
 print("No frame, reconnecting...")
 cap.release()
 time.sleep(1)
 cap = open_camera()
 continue

RTSP 流断开时自动重新连接，确保系统长时间稳定运行。

6 清理退出

释放摄像头 → 关闭 FFmpeg 管道 → 等待子进程结束 → 干净退出

FFmpeg 参数详解

参数	说明
`-f rawvideo -vcodec rawvideo -pix_fmt bgr24`	输入原始 BGR24 视频数据
`-s 640x480 -r 25`	输入分辨率与帧率
`-c:v libx264`	H.264 视频编码
`-preset ultrafast -tune zerolatency`	极速编码，零延迟
`-b:v 2500k -maxrate 2500k -bufsize 500k`	视频码率控制
`-f mpegts udp://IP:PORT?pkt_size=1316`	UDP 推流输出

VLC 播放地址

udp://@0.0.0.0:5600

Windows 端 VLC 媒体 → 打开网络串流 → 输入以上地址即可实时观看检测画面。

性能参考

Jetson Orin Nano (CUDA 加速)： ~25 FPS
YOLO 检测对象： 人 (class 0)
端到端延迟：< 200ms (相机 → 检测 → UDP 推流 → VLC 显示)

Related Drobotics guides